Neurobot Motion

Same motion, different score

The complexity index counts peaks in the power spectrum of a bot’s x and y positions over 30 minutes, above a fixed power. A steady circle has one. But peaks also come from things that aren’t new rhythms: noise in the beating, a square well’s corners, an orbit whose speed drifts, stops and starts. And the count depends on the window length and threshold, which the score fixes but the bot doesn’t care about.

Loading…

Would a study like theirs find this?

The study compared 46 neurobots with 48 biobots and found two differences: a higher complexity index and a higher minimum speed, with distance, mean speed, acceleration and area covered unchanged. Here each explanation is a change to a simulated bot, and each simulated study repeats theirs: the same numbers of bots, the same measures, the same tests. The bars show how often a study like theirs would report what they reported.

The biobots come in three versions, all calibrated to the paper’s published distributions: “stallers” stop now and then; “sitters” are a mix of bots that sit still all session and bots that move steadily, as in the paper’s own video; and the last also measures minimum speed the way the paper most likely did.

Loading the studies…

What we found

A flat body can’t reproduce both results

With biobots that stall, the two results pull against each other: stalls raise the complexity index and lower the minimum speed, so anything that raises one tends to lower the other. With biobots that either sit still or move steadily, they rise together, because waking a sitter raises both, but distance and mean speed rise with them, which the paper didn’t see. For every change to the flat, rigid body, all four of the paper’s criteria hold in fewer than one study in a hundred.

A rolling body comes closest, without neurons

Real bots roll in 3D, and neurobots are longer and about twice the area of biobots. A lopsided body rolling about its long axis moves its tracked center sideways once per turn, which adds spectral peaks without the bot ever stopping. Giving the flat body that wobble (0.2 mm per unit of elongation, which leaves the biobots’ published index distribution intact), neurobots that differ only in shape score about one point higher in two-thirds of studies of the paper’s size, the paper’s own difference. With fewer neurobots that sit still (the paper found 1 of 47 against 6 of 48) and their bigger bodies tracked a little more precisely, all four criteria hold in 1.6% of studies: the closest any explanation comes, and still rare.

Loading…

How minimum speed was probably measured

The paper doesn’t define minimum speed. Only one kind of reading reproduces its biobots’ minimum speeds and mean speeds together: the slowest single frame-to-frame speed, measured with tracking error about as large as a frame’s movement. At that slowest frame, tracking error can cancel a real step, so the reading depends on how precisely a bot is tracked as well as on how it moves. Smoothed readings, including the 10-second bins this project pre-registered, don’t fit.

Loading…

Or ordinary differences between batches

In a lab, two groups built from different clutches of embryos, on different days, or recorded in different sessions differ in ways nobody designed. Giving each group its own batch, with every body parameter off by a random 20% or so, and no neurons or designed difference at all, a study of the paper’s size finds the complexity index and minimum speed both higher in 11% of studies: more often than neural bursts, arrests or a rhythm do. It never matches all four results, because distance and speed shift too. Whether the paper’s groups were made and recorded side by side isn’t stated.

A reading without neurons, as a hypothesis

Three observations together give much of the paper’s result without neural control of the cilia. First, in the paper’s own supplementary data, sham bots (implanted with cells that don’t become neurons) had a complexity index distributed like the neurobots’, short of significance with 22 bots. Shams weren’t longer or bigger than biobots, but they did carry an implanted clump, and the paper tracked each bot’s center of mass in the video: a clump off center would make the tracked point wobble as the bot rolls, in shams and neurobots but not in biobots. Second, the bigger neurobots’ centers are probably tracked more precisely; tracking identical bots with 4 µm of error instead of 5 µm reproduces the paper’s minimum-speed medians with mean speed and the complexity index unchanged. Third, fewer neurobots sat still.

This rests on three things nobody has measured: how minimum speed was computed, how tracking error depends on bot size, and how much the tracked point wobbles as a bot rolls. Even together they match all four of the paper’s results only rarely, so chance may also play a part in a pair of results at p ≈ 0.04. The paper’s own explanations, neurons acting on cilia directly or through the cells around them, remain possible, and its other findings (neurons that grow and fire, a different transcriptome, different responses to a drug) are untouched. The tests that would decide it are cheap: re-analyze the raw videos for stops, speed variability and the tracked point’s wobble against the outline, and measure tracking error for bots of known size.

Score your own track

The same measures, for any track: paste or load positions over time and see the complexity index under several readings, three speeds, and warnings for the artifacts above.

Conclusion

The study found that neurobots move with a higher complexity index and a higher minimum speed than plain biobots, with distance, speed, acceleration and area unchanged, and read it as the neurons’ doing. Rebuilding the measurements and simulating studies of the same size suggests the two measures can’t carry that conclusion on their own. The complexity index counts irregularity, stops and the well’s geometry as well as rhythms, and shifts with analysis choices; the minimum speed was probably a single slowest frame, which depends on how precisely a bot is tracked. In a model of these bodies, no change, neural or not, reproduces all four results more than rarely. The closest is a bigger, lopsided body rolling as it swims, with no neurons involved, and ordinary differences between batches give the two headline results together about one study in nine. None of this shows the paper is wrong: the model is a simplification, and the study’s other findings (neurons that grow and fire, changed gene expression, a different drug response) stand. It does mean the behavioral result is open, and cheap to settle: re-analyze the raw videos for stops and wobble, and measure tracking error on bots of known size.

The science

The bots are a 2D model of a cilia-driven body: a rigid ellipse whose rim carries 64 sites, each a ciliated cell or not. Each cell pushes the body along a shared beat direction and around its rim, so bots swim in circles whose size depends on how the cells are spread. Beating has noise per cell and a slow shared drift; bots can stop, sit, or carry neural effects on the cells near an implant. Biobot populations were calibrated to the paper’s published distributions, and the main comparison was fixed in advance, before any neural change was simulated. The sitters, the rhythm, the minimum-speed reconstruction and the rolling wobble came after, and are labelled exploratory.

The findings come in two kinds. Those about the measures (what the complexity index counts, how the score moves with the window and threshold, how a slowest-frame speed depends on tracking) are arithmetic on positions over time, so they hold for the paper’s real tracks as much as for simulated ones; the paper’s own video showed the same artifacts. Those about mechanisms (which changes to a bot do or don’t reproduce the results) are statements about this model, and the physical experiment had more in it than the model does: batches of embryos and recording sessions, the tracking software and camera, light, temperature and handling, and which bots were recorded. Batch differences are simulated above; the tracker stands in only as tracking error and a wobble.